Fundemantals of Image Processing Course Project

Hypothesis = Each individual Gan type leaves a unique noise residual on each genarated image and also PRNU(Photo-Response-Non-Uniformly) pattern. We would like to see our results under this hypotesis.

Generate Style GAN's Pics

PRNU

Extract Fingerprints for Gan's

Extract Fingerprints for Normal's

Computing Residuals

Results

As seen from histograms, there are higher correlation between same classes.

Creating the heat map of correlations.

This correlation matrix shows the same results as the histogram of correlation above.

We defined that the best threshold value should be the intersection point of GAN and Normal pictures.

Classification

From 500 GAN and 500 Normal pictures we labeled them as 0 and 1 accordingly. Prediction column defined as randomly.

If we set threshold as 0, we find AUC as 0.732. However, we predicted 100% correctly of all the normal pictures.

Threshold tuning

As can be seen from the graph above, best threshold value is 0.005.

Test Picture

We computed fingerprints for test picture.

Firstly, we correlated test picture's fingerprint with 500 normal picture residual. Then we take mean of this normalized cross-correlation and see if it is higher or lower than the threshold.

Lasly, we selected a random StyleGAN2 picture with psi = 1.0 and threshold = 0,001, predicted as GAN picture, which is correct.

Tests

2) Gan Test Pic

3) Normal Pic Test

Our test pictures

According to the AUC graph we should've select the threshold as 0.005. However, we get better results when we select threshold as 0.001, so we stick to that value.

Test

Test with threshold = 0.005